The m-step least squares extrapolation is generally different from the m-step naive extrapolation in non-linear AR(1) models when n≥2. The authors show that there exists a class of non-linear AR(1) models in which a difference between these two extrapolations is arbitrary large.
